Aircraft simulation model confidence coefficient optimization method

Abstract

The invention relates to a method for optimizing the confidence coefficient of an aircraft simulation model, and provides a method for improving the confidence coefficient of a simulation model and solving the problem of data delay by combining a simulation system and aircraft transmission data in order to solve the problems that the confidence coefficient of the aircraft simulation model is not high and the online data transmission of an aircraft is delayed. Simulation output data corrected by the off-line network has the characteristic insufficiency phenomenon on real data, and the on-line network can carry out secondary correction on an off-line correction result by combining delayed transmission real data, so that the method disclosed by the invention combines two parts, namely an off-line correction network and an on-line correction network. A random forest algorithm and an MDA method are adopted to extract off-line network and on-line network features. The offline network adopts a bidirectional long short-term memory neural network (BiLSTM) network and a time attention mechanism to preliminarily correct simulation data errors; and carrying out secondary correction on the error between the offline correction result and the real data through an online algorithm FTRL.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft simulation digital twins, and in particular, relates to a method for optimizing the confidence of an aircraft simulation model. Background Art

[0002] With the rapid development of modern aviation technology, aircraft simulation models have been widely used in flight performance analysis, pilot training, aircraft simulation digital twins, and aircraft design. However, existing aircraft simulation models still have many problems in terms of accuracy, complexity, parameter uncertainty, and adaptability.

[0003] The current simulation model is still insufficient in terms of accuracy in simulating flight states and dynamic characteristics, resulting in a large deviation between the simulation results and the actual flight data, affecting its reliability. High-precision simulation models are usually complex in structure and large in computation, which makes it difficult to meet the needs of real-time simulation and rapid response, increasing development and maintenance costs. A large number of parameters in the model are difficult to accurately measure or estimate, resulting in large parameter uncertainty, which in turn affects the confidence of the simulation results. The existing simulation models have poor adaptability under different flight states and environmental conditions, especially in severe weather conditions, where the prediction accuracy is significantly reduced. In order to solve these problems, a confidence optimization method for aircraft simulation models is proposed. This method combines the simulation model input data, simulation model output data and real data, and uses the deep learning network BiLSTM and the temporal attention mechanism to preliminarily correct the error. Then, the error is corrected online by combining real-time real data and the FTRL algorithm, so that the optimization results are closer to the real data, thereby achieving the purpose of improving the confidence of the simulation model.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a confidence optimization method for an aircraft simulation model, thereby solving the problems raised in the above-mentioned background technology.

[0006] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0007] A confidence optimization method for an aircraft simulation model, comprising:

[0008] S1: Evaluation of simulation input features based on simulation model and MDA, evaluate the importance of simulation model input features to the simulation output features to be corrected, and retain high-importance simulation input features.

[0009] In view of the unchanged characteristics of simulation model parameters, the MDA baseline accuracy value of the simulation model is set to 0, and the simulation input features are shuffled in turn. The difference between the accuracy of the simulation output to be corrected and the accuracy of the shuffled simulation output to be corrected is calculated, the absolute value of the difference is taken, the absolute value is normalized, and the normalized result is used as the importance score. The simulation input features with high importance scores are saved as a high-importance input feature set.

[0010] The calculation formula for scrambling each simulation input feature is as follows:

[0011]

[0012] j is the jth shuffled simulation model input feature sequence, is the sample after the jth feature is randomly exchanged, n is the number of input samples, y i is the value output by the simulation model before disruption, To shuffle the jth feature, simulate the output of the model.

[0013] S2: Evaluation of simulation output features based on random forest MDA, through the simulation model output features, calculate the importance of simulation output features to the simulation output features to be corrected, and retain the output features with high importance;

[0014] Calculate the importance of simulation output features to the modified simulation output features; use the complete simulation model output features to train the random forest model and calculate the baseline accuracy of the model. For each output feature, shuffle the value of the feature in the data set and calculate the accuracy of the model on the shuffled data set. Subtract the absolute value of the baseline accuracy from the shuffled accuracy, and use the result after normalization of the absolute value as the importance score. Based on the score, evaluate the high-importance output feature set.

[0015] The features with higher importance scores are saved as high-importance output feature sets, and the calculation formula is:

[0016]

[0017] In the formula, the absolute value of MDA'(j) is taken as the importance evaluation score, and T is the number of random forest trees (X i ,y i ) is a sample, For X i The j-th dimension of the random exchange sample, D t is the out-of-bag sample set of random tree t, is the sample set R(X i ) is the sample X i The predicted output.

[0018] S3: Integrate high-importance input features, high-importance output features, and simulation output features to be corrected into a simulation data enhanced feature set, use the simulation data enhanced feature set as the input of a bidirectional long- and short-term offline correction network with a temporal attention mechanism, and use the difference between the real data and the simulation output data to be corrected as a label for training and testing;

[0019] The simulation data enhanced feature set is divided into a training set and a test set with a ratio of 0.8:0.2. The difference between the real data and the simulation output data is used as the label, and the error of offline correction plus the simulation model output is used as the result of offline correction.

[0020] S4: Use the BiLSTM network and the temporal attention mechanism algorithm for training, save the trained BiLSTM network, use the trained offline network to make preliminary corrections to the errors, and use the corrected errors plus the simulation results to be corrected as the final results of the offline correction;

[0021] The BiLSTM network and the time attention mechanism algorithm are used to correct the error between the true value and the simulated output value to be corrected, so that the result after adding the corrected error to the simulated output is close to the value of the aircraft at the real moment. BiLSTM is composed of two independent LSTMs, and the LSTM unit is composed of an input gate, a forget gate, an output gate, and a cell state. The main calculation formula of the LSTM unit is as follows:

[0022] Input gate: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i );

[0023] Forget gate: f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f );

[0024] Cell Status:

[0025] Cell status update:

[0026] Output gate: o t =σ(W xo x t +W ho h t-1 +W co ct +b o );

[0027] Hide status update:h t =o t ·tanh(c t );

[0028] In the formula, x t is the input at time step t, h t-1 is the hidden state of the previous time step, c t-1 is the cell state at the previous time step, i t is the input gate, f t For the Gate of Forgetfulness, is the cell state, c t Update the state, W and b are weight parameters.

[0029] In BiLSTM, the input data of each time step will be processed by two independent LSTM processors, one from front to back and the other from back to front. The outputs of the forward LSTM layer and the reverse LSTM layer are connected to form a comprehensive output vector. The temporal attention mechanism calculates the attention score of each time step on the output of the bidirectional LSTM and generates a context vector. The context vector is combined with the output of the bidirectional LSTM and passed to the fully connected layer to generate the final corrected output. The attention calculation steps are as follows:

[0030] The forward LSTM processes the sequence from the beginning to the end, and the backward LSTM processes the sequence from the end to the beginning. The final BiLSTM output is the connection of the forward and backward states. The formula is as follows:

[0031]

[0032] Calculate the attention score: Calculate the attention score of each time step through linear transformation, the formula is as follows:

[0033] e t =W att ·H t ;

[0034] Normalized attention score: Use the softmax function to normalize the score to weight, the formula is as follows:

[0035]

[0036] Calculate the context vector: Use the attention weights to weight the BiLSTM outputs to get the context vector. The formula is as follows:

[0037]

[0038] The context vector c is finally corrected and output through the fully connected layer. The formula is as follows:

[0039] output=W fc ·c+b fc .

[0040] Dropout is set in each LSTM hidden layer to make the activation value of a neuron stop working with a certain probability p, making the model more generalizable. The BiLSTM network uses the MSE loss function and Adam optimizer. During the forward propagation of the BiLSTM network, the input data passes through the BiLSTM and time attention mechanism to generate values. The BiLSTM network uses the MSE loss function to calculate the correction error during training. The loss between the actual error y is as follows:

[0041]

[0042] Calculate the gradient of loss L with respect to model parameters through back propagation Using the Adam optimizer, according to the gradient Update the model parameter θ, the calculation formula is as follows:

[0043] m t =β1m t-1 +(1-β1)g t ;

[0044]

[0045] In the formula, L is the loss value, which represents the loss value between the corrected error value and the true error value, N is the number of samples, the total number of samples in the data set, and y is i is the true error value of the i-th sample, is the model correction error value of the i-th sample, m t is the first-order momentum estimate at time step t, representing the exponentially weighted moving average of the gradient, v t is the second-order momentum estimate at time step t, which represents the exponentially weighted moving average of the squared gradient, β1 is the first-order momentum decay factor, β2 is the second-order momentum decay factor, and g t is the gradient of time step t, which represents the gradient of the loss function with respect to the parameters, θ t is the model parameter at time step t, θ t+1 is the model parameter at time step t+1, α is the learning rate, which controls the step size of parameter update, and ∈ is a small constant to prevent the denominator from being zero.

[0046] S5: Evaluate the real features in the real data set that are important to the simulated output features to be corrected, and retain the real feature set with high importance;

[0047] S6: Online network training and correction based on the FTRL algorithm. The high-importance real feature set and the offline output final result are reconstructed into a time series. The time series is used as the online network FTRL training input. The difference between the real feature data corresponding to the feature to be corrected and the offline output final result is used as the online training label. The error of online training plus the offline final result is used as the final result of network optimization. The evaluated high-importance real feature set is input to the algorithm FTRL for correction. The difference between the real feature data and the offline output final result is used as the online training label. A time sliding window of size 10 is set for online training and correction. The correction error of online training plus the result of offline correction is used as the final result of the entire network optimization.

[0048] The FTRL algorithm adopts different update methods for the weights of online input features. The weight of the i-th dimension feature is w i The gradient vector parameter is g i , the weight update steps are as follows:

[0049] The cumulative gradient update formula is as follows:

[0050]

[0051] Cumulative sum of squared gradients n i The update formula is as follows:

[0052]

[0053] The weight w for the feature i The update formula is as follows:

[0054]

[0055] Where α and β are hyperparameters, λ1>0, λ2>0 are L1 and L2 regularization coefficients respectively. i is the cumulative gradient of feature i, g i is the gradient of feature i in the previous step, n i is the cumulative gradient sum of feature i, w i is the model weight corresponding to feature i.

[0056] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0057] 1. The difference between the real value and the output value of the simulation model is preliminarily optimized through the offline network BiLSTM, and then the error between the real data and the offline network optimization result is secondary optimized with the real online data. The optimized error plus the final result of the offline network optimization is used as the final optimization output, which improves the confidence of the simulation system and makes the simulation output closer to the real data.

[0058] 2. Solve the problem of data transmission delay in the "cloud box" of future aircraft development, and ensure the authenticity and timeliness of simulation data.

[0059] 3. Importance assessment combines the characteristics of simulation models and simulation input data to significantly improve the accuracy of feature selection.

[0060] 4. The use of BiLSTM and temporal attention mechanism enhances the model's ability to process time series data and effectively captures the temporal correlation in the data.

[0061] 5. Through the Dropout and Adam optimization algorithms, the training effect and stability of the network are improved, and the risk of overfitting is reduced.

[0062] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0064] In the figure:

[0065] Figure 1 The figure is a schematic diagram of the method involved.

[0066] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0067] The present invention will now be described in further detail with reference to the accompanying drawings.

[0068] See also Figure 1 As shown, in the present implementation of simulation output pitch angle correction, a method for optimizing the confidence of an aircraft simulation model is provided, comprising:

[0069] Obtain input data (such as flight control parameters, environmental conditions, etc.) and output data (such as flight attitude, speed, position, etc.) from the aircraft simulation model; obtain corresponding output data from actual flight records, and perform data cleaning and normalization on the collected simulation data and real flight data;

[0070] S1: Evaluation of simulation input features based on simulation model and MDA, evaluate the importance of simulation model input features to the corrected simulation output pitch angle, and retain high-importance simulation input features (captain's side joystick force sensor signal, center of gravity estimation, left inner elevator actuator position, right elevator position after voting, right aileron position after voting, crosswind calculation, co-pilot's side joystick force sensor signal, center of gravity position at takeoff, left outer elevator actuator position, flaps, total fuel consumption during flight, turbulence, captain's side pedal force sensor signal, engine thrust rating, left elevator position after voting, cockpit-flap configuration position, total fuel, static air temperature, co-pilot's side pedal force sensor signal, right inner elevator actuator position, flap position (corrected), wind speed, engine throttle lever angle (corrected position), center of gravity position, elevator sensing position (right RVDT) signal, right outer elevator actuator position (LVDT) signal, left aileron position after voting, wind direction);

[0071] Combined with the unchanged characteristics of simulation model parameters, the MDA baseline accuracy value of the simulation model is set to 0, and the simulation input features are shuffled in turn. The difference between the simulation output pitch angle and the shuffled pitch angle is calculated, the absolute value of the difference is taken, the absolute value is normalized, and the normalized result is used as the importance score.

[0072] The calculation formula for scrambling each simulation input feature is as follows:

[0073]

[0074] j is the jth shuffled simulation model input feature sequence, is the sample after the jth feature is randomly exchanged, n is the number of input samples, y i is the value output by the simulation model before disruption, To shuffle the jth feature, simulate the output of the model.

[0075] S2: Evaluation of simulation output features based on random forest MDA, through the simulation model output features, calculate the importance of simulation output features to the simulated output pitch angle to be corrected, and retain high-importance output features (altitude change rate data, fuel-oil heat exchanger pressure difference, right inertial reference unit hybrid GPS altitude, left inertial reference unit hybrid GPS altitude, calculated airspeed from integrated standby flight display (ISFD));

[0076] Calculate the importance of simulation output features to the corrected simulation output pitch angle; use the output features of the complete simulation model to train the random forest model and calculate the baseline accuracy of the model. For each output feature, shuffle the value of the feature in the data set and calculate the accuracy of the model on the shuffled data set. Subtract the absolute value of the baseline accuracy from the shuffled accuracy, and take the result after normalization of the absolute value as the importance score. According to the score, evaluate the high-importance output feature set.

[0077] The features with higher importance scores are saved as high-importance output feature sets, and the calculation formula is:

[0078]

[0079] In the formula, the absolute value of MDA'(j) is taken as the importance evaluation score, and T is the number of random forest trees (X i ,y i ) is a sample, For X i The j-th dimension of the random exchange sample, D t is the out-of-bag sample set of random tree t, is the sample set R(X i ) is the sample X i The predicted output.

[0080] S3: Integrate high-importance input features, high-importance output features and the simulated output pitch angle to be corrected into a simulation data enhanced feature set, use the simulation data enhanced feature set as the input of the bidirectional long-term and short-term offline correction network of the temporal attention mechanism, and use the difference between the real pitch angle data and the simulated output pitch angle data to be corrected as a label for training and testing;

[0081] The simulation data enhanced feature set is divided into a training set and a test set with a ratio of 0.8:0.2. The difference between the true pitch angle and the simulation output pitch angle is used as the label, and the offline corrected error plus the simulation model output pitch angle is used as the offline corrected pitch angle.

[0082] S4: BiLSTM network and temporal attention mechanism algorithm are used for training and preliminary error correction, and the simulation data enhanced feature set is used as training data input, and the difference between the actual pitch angle and the simulated output pitch angle to be corrected is used as the label;

[0083] The BiLSTM network and the time attention mechanism algorithm are used to correct the error between the actual pitch angle and the simulated output pitch angle, so that the result after adding the corrected error to the simulated output is close to the pitch angle value corresponding to the actual moment of the aircraft. The BiLSTM is composed of two independent LSTMs, and the LSTM unit is composed of an input gate, a forget gate, an output gate, and a cell state. The main calculation formula of the LSTM unit is as follows:

[0084] Input gate: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i );

[0085] Forget gate: f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f );

[0086] Cell Status:

[0087] Cell status update:

[0088] Output gate: o t =σ(W xo x t +W ho h t-1 +W co c t +b o );

[0089] Hide status update:h t =o t ·tanh(c t );

[0090] In the formula, x t is the input at time step t, h t-1 is the hidden state of the previous time step, c t-1 is the cell state at the previous time step, i t is the input gate, f t For the Gate of Forgetfulness, is the cell state, c t Update the state, W and b are weight parameters.

[0091] In BiLSTM, the input data of each time step will be processed by two independent LSTM processors, one from front to back and the other from back to front. The outputs of the forward LSTM layer and the reverse LSTM layer are connected to form a comprehensive output vector. The temporal attention mechanism calculates the attention score of each time step on the output of the bidirectional LSTM and generates a context vector. The context vector is combined with the output of the bidirectional LSTM and passed to the fully connected layer to generate the final corrected output. The attention calculation steps are as follows:

[0092] The forward LSTM processes the sequence from the beginning to the end, and the backward LSTM processes the sequence from the end to the beginning. The final BiLSTM output is the connection of the forward and backward hidden states. The formula is as follows:

[0093]

[0094] Calculate the attention score: Calculate the attention score of each time step through linear transformation, the formula is as follows:

[0095] e t =W att ·H t ;

[0096] Normalized attention score: Use the softmax function to normalize the score to weight, the formula is as follows:

[0097]

[0098] Calculate the context vector: Use the attention weights to weight the BiLSTM outputs to get the context vector. The formula is as follows:

[0099]

[0100] The context vector c is finally corrected and output through the fully connected layer. The formula is as follows:

[0101] output=W fc ·c+b fc .

[0102] Dropout is set in each LSTM hidden layer to make the activation value of a neuron stop working with a certain probability p, making the model more generalizable. The BiLSTM network uses the MSE loss function and Adam optimizer. During the forward propagation of the BiLSTM network, the input data passes through the BiLSTM and time attention mechanism to generate values. The BiLSTM network uses the MSE loss function to calculate the correction error during the training process The loss between the actual error y is as follows:

[0103]

[0104] Calculate the gradient of loss L with respect to model parameters through back propagation Using the Adam optimizer, according to the gradient Update the model parameter θ, the calculation formula is as follows:

[0105] m t =β1m t-1 +(1-β1)g t ;

[0106]

[0107] In the formula, L is the loss value, which represents the loss value between the corrected error value and the true error value, N is the number of samples, the total number of samples in the data set, and y is i is the true error value of the i-th sample, is the model correction error value of the i-th sample, m t is the first-order momentum estimate at time step t, representing the exponentially weighted moving average of the gradient, v t is the second-order momentum estimate at time step t, which represents the exponentially weighted moving average of the squared gradient, β1 is the first-order momentum decay factor, β2 is the second-order momentum decay factor, and g t is the gradient of time step t, which represents the gradient of the loss function with respect to the parameters, θ t is the model parameter at time step t, θ t+1 is the model parameter at time step t+1, α is the learning rate, which controls the step size of parameter update, and ∈ is a small constant to prevent the denominator from being zero.

[0108] S5: Evaluate the importance of the features in the real data set to the corrected simulation output pitch angle, and retain the high-importance real feature set (altitude change rate data, fuel-oil heat exchanger pressure difference, right inertial reference unit hybrid GPS altitude, left inertial reference unit hybrid GPS altitude, calculated airspeed from the integrated standby flight display (ISFD));

[0109] S6: Online network training and correction based on FTRL algorithm, reconstructing the time series of high-importance real feature sets and offline output pitch angles, taking the time series as the input of online network FTRL training, the difference between the real pitch angle and the final offline network corrected pitch angle is taken as the online training label, and the error of online training plus the final result of offline network is taken as the final pitch angle of network optimization;

[0110] The highly important real feature set is evaluated and input to the online algorithm FTRL for correction. A time sliding window of size 10 is set for online training and correction. The correction error of online training plus the pitch angle of the offline network is used as the result of the optimized pitch angle of the entire network.

[0111] The FTRL algorithm adopts different update methods for the weights of online input features. The weight of the i-th dimension feature is w i The gradient vector parameter is g i , the weight update steps are as follows:

[0112] The cumulative gradient update formula is as follows:

[0113] The cumulative gradient update formula is as follows:

[0114]

[0115] Cumulative sum of squared gradients n i The update formula is as follows:

[0116]

[0117] The weight w for the feature i The update formula is as follows:

[0118]

[0119] Where α and β are hyperparameters, λ1>0, λ2>0 are L1 and L2 regularization coefficients respectively. i is the cumulative gradient of feature i, g i is the gradient of feature i in the previous step, n i is the cumulative gradient sum of feature i, w i is the model weight corresponding to feature i.

[0120] The offline network BiLSTM is used to preliminarily optimize the difference between the actual pitch angle and the pitch angle output by the simulation model, and then the error between the actual data and the offline network optimization result is secondary optimized with the real online data. The optimized error plus the final result of the offline network optimization is used as the final optimization output, which improves the confidence of the simulation system and makes the simulation output closer to the real data. It solves the data transmission delay problem of the "cloud box" in future aircraft research and development, and the authenticity and timeliness of the simulation data. The importance assessment combines the characteristics of the simulation model and the simulation input data to significantly improve the feature selection accuracy. The BiLSTM and temporal attention mechanism are used to enhance the model's ability to process time series data and effectively capture the temporal correlation in the data. The Dropout and Adam optimization algorithms are used to improve the training effect and stability of the network and reduce the risk of overfitting.

[0121] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A confidence optimization method for an aircraft simulation model, characterized in that: include: S1: Evaluation of simulation input features based on simulation model and MDA, evaluating the importance of simulation model input features to the simulation output features to be corrected, and retaining high-importance simulation input features; S2: Evaluation of simulation output features based on random forest MDA, through the simulation model output features, calculate the importance of simulation output features to the simulation output features to be corrected, and retain the output features with high importance; S3: Integrate high-importance input features, high-importance output features, and simulation output features to be corrected into a simulation data enhanced feature set, use the simulation data enhanced feature set as the input of a bidirectional long- and short-term offline correction network with a temporal attention mechanism, and use the difference between the real data and the simulation output data to be corrected as a label for training and testing; S4: Use the BiLSTM network and the temporal attention mechanism algorithm for training, save the trained BiLSTM network, use the trained offline network to make preliminary corrections to the errors, and use the corrected errors plus the simulation results to be corrected as the final results of the offline corrections; S5: Evaluate the real feature data in the real data set that are important to the features of the simulated output to be corrected, and retain them as a high-importance real feature set; S6: Online network training and correction based on the FTRL algorithm. The high-importance real feature set and the offline output final result are reconstructed into a time series. The time series is used as the online network FTRL training input. The difference between the real feature data corresponding to the feature to be corrected and the offline output final result is used as the online training label. The error of online training plus the offline final result is used as the final result of network optimization.

2. The method for optimizing the confidence of an aircraft simulation model according to claim 1, characterized in that: In step S1, in combination with the unchanged characteristics of the simulation model parameters, the MDA baseline accuracy value of the simulation model is set to 0, and the simulation input features are shuffled in turn. The difference between the accuracy of the simulation output to be corrected and the accuracy of the shuffled simulation output to be corrected is calculated, the absolute value of the difference is taken, the absolute value is normalized, and the normalized result is used as the importance score. The simulation input features with high importance scores are saved as a high-importance input feature set.

3. The aircraft simulation model confidence optimization method according to claim 2, characterized in that: The calculation formula for scrambling each simulation input feature is as follows: j is the jth shuffled simulation model input feature sequence, is the sample after the jth feature is randomly exchanged, n is the number of input samples, y i is the value output by the simulation model before disruption, To shuffle the jth feature, simulate the output of the model.

4. The method for optimizing the confidence of an aircraft simulation model according to claim 1, characterized in that: In step S2, the importance of the simulation output features to the modified simulation output features is calculated by simulating the model output features, a random forest model is trained using the complete simulation model output features, and the baseline accuracy of the model is calculated. For each output feature, the value of the feature in the data set is shuffled, and the accuracy of the model on the shuffled data set is calculated. The absolute value of the baseline accuracy minus the shuffled accuracy is taken, and the result after normalization of the absolute value is used as the importance score. According to the score, a set of high-importance output features is evaluated.

5. The aircraft simulation model confidence optimization method according to claim 4, characterized in that: The features with higher importance scores are saved as high-importance output feature sets, and the calculation formula is: In the formula, the absolute value of MDA'(j) is taken as the importance evaluation score, and T is the number of random forest trees (X i ,y i ) is a sample, For X i The j-th dimension of the random exchange sample, D t is the out-of-bag sample set of random tree t, is the sample set R(X i ) is the sample X i The predicted output.

6. The aircraft simulation model confidence optimization method according to claim 1, characterized in that: In step S3, the simulation data enhanced feature set is used as the input of the BiLSTM network, the difference between the real data and the simulation output data is used as the label, and the offline corrected error plus the simulation model output data is used as the final result of offline correction.

7. The aircraft simulation model confidence optimization method according to claim 1, characterized in that: In step S4, the BiLSTM network and the temporal attention mechanism algorithm are used to correct the error between the true value and the simulated output value, so that the result after adding the corrected error to the simulated output is close to the true value of the aircraft. The BiLSTM is composed of two independent LSTMs, and the LSTM unit is composed of an input gate, a forget gate, an output gate, and a cell state. The main calculation formula of the LSTM unit is as follows: Input gate: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ); Forget gate: f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f ); Cell Status: Cell status update: Output gate: o t =σ(W xo x t +W ho h t-1 +W co c t +b o ); Hide status update:h t =o t ·tanh(c t ); In the formula, x t is the input at time step t, h t-1 is the hidden state of the previous time step, c t-1 is the cell state at the previous time step, i t is the input gate, f t For the Gate of Forgetfulness, is the cell state, c t Update the state, W and b are weight parameters.

8. The aircraft simulation model confidence optimization method according to claim 7, characterized in that: In BiLSTM, the input data of each time step will be processed by two independent LSTM processors, one from front to back and the other from back to front. The outputs of the forward LSTM layer and the reverse LSTM layer are connected to form a comprehensive output vector. The temporal attention mechanism calculates the attention score of each time step on the output of the bidirectional LSTM and generates a context vector. The context vector is combined with the output of the bidirectional LSTM and passed to the fully connected layer to generate the final corrected output. The attention calculation steps are as follows: The forward LSTM processes the sequence from the beginning to the end, and the backward LSTM processes the sequence from the end to the beginning. The final BiLSTM output is the connection of the forward and backward hidden states. The formula is as follows: Calculate the attention score: Calculate the attention score of each time step through linear transformation, the formula is as follows: have been t =W att H t ; Normalized attention score: Use the softmax function to normalize the score to weight, the formula is as follows: Calculate the context vector: Use the attention weights to weight the BiLSTM outputs to get the context vector. The formula is as follows: The context vector c is finally corrected and output through the fully connected layer. The formula is as follows: output=W fc ·c+b fc 。 9. The aircraft simulation model confidence optimization method according to claim 8, characterized in that: Dropout is set in each LSTM hidden layer to make the activation value of a neuron stop working with a certain probability p, making the model more generalizable. The BiLSTM network uses the MSE loss function and Adam optimizer. During the forward propagation of the BiLSTM network, the input data passes through the BiLSTM and time attention mechanism to generate values. The BiLSTM network uses the MSE loss function to calculate the correction error during training. The loss between the actual error y is as follows: Calculate the gradient of loss L with respect to model parameters through back propagation Using the Adam optimizer, according to the gradient Update the model parameter θ, the calculation formula is as follows: m t =β1m t-1 +(1-β1)g t ; In the formula, L is the loss value, which represents the loss value between the corrected error value and the true error value, N is the number of samples, the total number of samples in the data set, and y is i is the true error value of the i-th sample, is the model correction error value of the i-th sample, m t is the first-order momentum estimate at time step t, representing the exponentially weighted moving average of the gradient, v t is the second-order momentum estimate at time step t, which represents the exponentially weighted moving average of the squared gradient, β1 is the first-order momentum decay factor, β2 is the second-order momentum decay factor, and g t is the gradient of time step t, which represents the gradient of the loss function with respect to the parameters, θ t is the model parameter at time step t, θ t+1 is the model parameter at time step t+1, α is the learning rate, which controls the step size of parameter update, and ∈ is a small constant to prevent the denominator from being zero.

10. The aircraft simulation model confidence optimization method according to claim 1, characterized in that: In step S6, the high-importance real feature set is input to the algorithm FTRL for correction. The difference between the real feature data and the offline output final result is used as the online training label. A time sliding window of size 10 is set for online training and correction. The correction error of the online training plus the offline final result is used as the final result of the entire network optimization. The FTRL algorithm adopts different update methods for the weights of online input features. The weight of the i-th dimension feature is w i The gradient vector parameter is g i , the weight update steps are as follows: The cumulative gradient update formula is as follows: Cumulative sum of squared gradients n i The update formula is as follows: The weight w for the feature i The update formula is as follows: Where α and β are hyperparameters, λ1>0, λ2>0 are L1 and L2 regularization coefficients respectively. i is the cumulative gradient of feature i, g i is the gradient of feature i in the previous step, n i is the cumulative gradient sum of feature i, w i is the model weight corresponding to feature i.

Images